【作者】 陈睿；

【导师】 李华；

【作者基本信息】 中国科学院研究生院（计算技术研究所） ， 计算机应用技术， 2005， 博士

【摘要】 本文的主要研究内容是从多个同步的视频序列中自动恢复人体的三维运动姿态。这种无标记的人体运动捕捉跟踪技术可广泛应用于体育运动分析、医学诊断、虚拟现实、计算机动画、视频监控、人机交互等领域。由于存在非刚体人体描述、人体模型的三维到二维投影多义性、人体模型的自遮挡、高维状态空间搜索、复杂条件下的图像特征提取与匹配等方面的困难,从视频图像中恢复出人体三维运动姿态存在大量的不确定性。因此三维人体运动跟踪是计算机视觉领域一项非常有挑战性的任务。本文提出了有模型指导的三维人体运动跟踪框架,将一个多关节的圆台形状三维人体模型与多个视频图像中的外轮廓、边界、灰度和肤色特征进行匹配,使人体运动跟踪变成一个状态估计问题。并且,使用基于概率模型的粒子滤波算法来完成非线性、非高斯动态系统的状态估计。粒子滤波算法虽然能在混乱背景及遮挡情况下很好地完成一般跟踪任务,但是对于人体运动估计仍然存在困难。因此本文提出了两种新的粒子滤波改进策略。一种是将状态空间分解和PERM(Pruned-Enriched Rosenbluth Method)采样与退火粒子滤波结合,提高了对多模式后验分布的模拟精度。另一种是确定性搜索方法与随机采样方法相结合的改进粒子滤波算法,用于跟踪复杂背景下的三维人体运动。这种新的粒子滤波算法的最大特点就是通过局部优化方法来指导重要性分布函数的生成,使得对高维空间中多峰后验分布函数的估计成为可能。此外,在图像特征提取方面,本文提出了一种非参数的背景估计模型,用于检测视频图像中的运动人体的外轮廓。这种方法综合考虑了图像上的时间与空间信息,利用颜色和边界特征增强前景检测的可靠性;而且,通过自适应的阴影消除,进一步提高了运动目标检测的准确性。本文算法在模拟和真实数据上进行了试验,能够完成复杂背景条件下的人体运动跟踪任务。更多还原

【Abstract】 This thesis focuses on the automatic recovery of three-dimensional human motion from multiple synchronized video sequences. The potential applications of this kind of markerless motion capture technique are motion analysis, medical diagnosis, virtual reality, computer animation, video surveillance, human-computer interface and so on. 3D human motion tracking faces difficulties caused by non-rigid human model representation, 2D-3D projection, self occlusion, high dimensionality of state space and image features extraction under clutter. It is a challenging task in the field of computer vision.This thesis proposes a model based 3D human motion tracking framework, where the articulated human model represented by truncated cones is matched with several image features, such as silhouette, edge, intensity and skin color. And human motion tracking is the problem of state estimation, which can be accomplished by the particle filter algorithm based on probabilistic model.The particle filter algorithm, having the advantage of tracking under clutter and self occlusion, still suffers the pain of high computing complexity during 3D human motion estimation. This thesis proposes two improvements of standard particle filter. Firstly, for the purpose of improving the accuracy of posterior distribution, state space decomposition and PERM (Pruned-Enriched Rosenbluth Method) sampling are adopted during the annealed particle filter. Secondly, a new particle filter based sampling framework, which combines the local optimization and stochastic sampling, is proposed. The most important feature of this sampling method is that the optimization result is used to guide the importance function, which suits for estimation of multi-modal distribution in high dimensionality.Further more, this thesis proposes a novel method based on a non-parametric background model to detect the silhouette of human body in video sequences. This background subtraction method utilizes intensity and edge features synchronously to improve robustness of the foreground detection. And an adaptive shadow detection model is used to find the accurate moving objects.The algorithm is tested on simulative and real video sequences, which include human motion with self-occlusion, and can accomplish the 3D human motion tracking tasks.更多还原